sulus-web-mcp
This server provides stateless browser-based web research via three Playwright MCP tools, each opening a URL, extracting content, and closing the browser context.
browse_page: Fetch a public page's title, final URL, and cleaned markdown (or links/title only) with configurable wait and extraction modes.
extract_links: Get a markdown list of unique outbound links (href + text), optionally limited and restricted to the same origin.
search_page: Search a page's text for a case-insensitive query and return up to N snippets with surrounding context.
Each call is self-contained: navigate, extract, close; no cookies or sessions shared between calls.
No click, type, login, or file-upload capabilities.
Enforces safety by blocking localhost, private networks, and other risky hosts; HTTPS only in production.
Usable over HTTP MCP with bearer auth or stdio for local Cursor, and ready for ai-phone-system agents.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@sulus-web-mcp Browse https://example.com and summarize the content"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
sulus-web-mcp
Stateless Playwright MCP for live-web research. Agents get three tools — browse_page, extract_links, search_page — each of which opens a URL, extracts text, and closes the browser context. No click/type/login tools.
Built for Sulus Agents (ai-phone-system MCP connectors) and the rest of the ecosystem (Cursor, MCPConnect). Laravel cannot run Chromium in-process; this service is the HTTP MCP those agents call.
Tools
Tool | When to use | What it returns |
| Need the content of a public page | Title, final URL, cleaned markdown (or links/title only) |
| Need outbound links | Markdown list of href + text |
| Need excerpts matching a query | Up to N snippets around the query |
Each call is self-contained (navigate + extract). That matches ai-phone-system's connector client, which POSTs tools/list / tools/call with no MCP session.
Related MCP server: mcp-playwright
Requirements
Node.js 20+
Playwright Chromium (
npx playwright install chromiumafternpm install)
Setup
cd /Users/macbook/Workshop/Sulus.ai/sulus-web-mcp
npm install
npx playwright install chromium
cp .env.example .env
# set MCP_SHARED_SECRET
npm run buildHTTP (Agents, team, production)
npm run start:http
# or: npm run dev:httpHealth: GET /health
MCP: POST /mcp (JSON-RPC, Streamable HTTP)
Variable | Description |
| Bearer token required on |
| Bind (default |
| Allowed |
| Set |
| Optional host lists |
| Default 25000 (stay under Agents' 60s tool timeout) |
| Default 3 |
| Per bearer token, default 30 |
| Default 18000 (Agents truncate at 20k) |
Logs are one JSON line per browse (url origin+path, status, ms). Page bodies are never logged.
Production (browse.sulus.ai, same droplet as MiniERP)
Native Node + systemd + nginx — no Docker. Runbook, unit file, nginx, and ./deploy.sh:
# On the droplet, as deploy:
cp deploy/vps/env.production.example .env.production # set MCP_SHARED_SECRET
./deploy.shstdio (local Cursor)
npm run start.cursor/mcp.json:
{
"mcpServers": {
"sulus-web": {
"command": "node",
"args": ["dist/index.js"],
"cwd": "/Users/macbook/Workshop/Sulus.ai/sulus-web-mcp"
}
}
}HTTP from Cursor (after deploy):
{
"mcpServers": {
"sulus-web": {
"url": "https://browse.sulus.ai/mcp",
"headers": {
"Authorization": "Bearer YOUR_SHARED_SECRET"
}
}
}
}Connect from ai-phone-system Agents
Deploy this service behind HTTPS.
In MCP Connectors, add a custom server:
URL:
https://browse.sulus.ai/mcpAuth: Bearer, token =
MCP_SHARED_SECRET
Enable the connector on the agent.
The HTTP handler answers tools/list and tools/call without initialize / Mcp-Session-Id, which is what ExternalMcpClientService sends. Sessionful Streamable HTTP still works for Cursor.
Safety
HTTPS only (unless
ALLOW_INSECURE_HTTP=true)Blocks localhost, RFC1918, link-local, CGNAT, cloud metadata hosts
Resolves DNS and rejects private answers
Aborts in-page requests to blocked URLs
New browser context per call — no cookie sharing
No form-fill / click / file-upload tools
Scripts
npm run typecheck
npm test
npm run buildAvailable Tools
3 toolsbrowse_pageA
Use this when you need the content of a live public web page. It opens the URL in a headless browser, waits for the document, and returns the title, final URL, and cleaned markdown (or links/title only). Returns extracted text, not a screenshot. Does not log in or fill forms.
Inputs: url (required), wait, extract.
extract_linksA
Use this when you need outbound links from a public page. It opens the URL and returns unique hrefs with link text. Returns a markdown list. Optionally restrict to the same origin.
Inputs: url (required), limit, same_origin.
search_pageA
Use this when you need excerpts matching a query on a public page. It opens the URL, extracts text, and returns up to N case-insensitive snippets around the query. Returns markdown snippets.
Inputs: url (required), query (required), max_matches, context_chars.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 3 tools
Each tool has a clearly distinct purpose: browse_page returns page content, extract_links retrieves outbound links, and search_page provides matching snippets. No ambiguity exists between them despite all operating on URLs.
All tool names follow a consistent verb_noun pattern (browse_page, extract_links, search_page) using snake_case. The pattern is predictable and uniform across the three tools.
Three tools is a well-scoped count for a web browsing server. Each tool handles a distinct aspect (content, links, search) and none are redundant or missing for the core functionality.
The tool surface covers the primary browsing capabilities (content extraction, link discovery, and targeted searching). Minor gaps like screenshot capture or form interaction exist, but they are beyond the stated scope and the tools together handle the typical use cases effectively.
Related MCP Connectors
Live web access for agents: scrape, SERP search, crawl/map, 100+ collectors, datasets, proxies.
Hosted browser for agents: open pages, search the web, extract data, read PDFs and take screenshots.
Agent-native search engine with live web research optimized for AI agents.
Web scraping, search, crawling, structured extraction and research for AI agents.
Related MCP Servers
- AlicenseNot gradedqualityFmaintenanceEnables web search across Bing, DuckDuckGo, and Google using Playwright automation, with tools for retrieving search results and extracting page text content.12 npmMIT
- AlicenseBqualityDmaintenanceProvides browser automation capabilities for LLM applications, enabling web page interaction and data extraction via Playwright.1246 PyPI10MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to control a web browser using Playwright, supporting navigation, interaction, and data extraction through natural language.MIT
- AlicenseAqualityDmaintenanceEnables web search, content fetching, dynamic page scraping, and URL preview via DuckDuckGo, Playwright, and Cheerio, without requiring API keys.4ISC